1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Support for Partition Mobility/Migration
4  *
5  * Copyright (C) 2010 Nathan Fontenot
6  * Copyright (C) 2010 IBM Corporation
7  */
8 
9 
10 #define pr_fmt(fmt) "mobility: " fmt
11 
12 #include <linux/cpu.h>
13 #include <linux/kernel.h>
14 #include <linux/kobject.h>
15 #include <linux/nmi.h>
16 #include <linux/sched.h>
17 #include <linux/smp.h>
18 #include <linux/stat.h>
19 #include <linux/stop_machine.h>
20 #include <linux/completion.h>
21 #include <linux/device.h>
22 #include <linux/delay.h>
23 #include <linux/slab.h>
24 #include <linux/stringify.h>
25 
26 #include <asm/machdep.h>
27 #include <asm/rtas.h>
28 #include "pseries.h"
29 #include "../../kernel/cacheinfo.h"
30 
31 static struct kobject *mobility_kobj;
32 
33 struct update_props_workarea {
34 	__be32 phandle;
35 	__be32 state;
36 	__be64 reserved;
37 	__be32 nprops;
38 } __packed;
39 
40 #define NODE_ACTION_MASK	0xff000000
41 #define NODE_COUNT_MASK		0x00ffffff
42 
43 #define DELETE_DT_NODE	0x01000000
44 #define UPDATE_DT_NODE	0x02000000
45 #define ADD_DT_NODE	0x03000000
46 
47 #define MIGRATION_SCOPE	(1)
48 #define PRRN_SCOPE -2
49 
50 static int mobility_rtas_call(int token, char *buf, s32 scope)
51 {
52 	int rc;
53 
54 	spin_lock(&rtas_data_buf_lock);
55 
56 	memcpy(rtas_data_buf, buf, RTAS_DATA_BUF_SIZE);
57 	rc = rtas_call(token, 2, 1, NULL, rtas_data_buf, scope);
58 	memcpy(buf, rtas_data_buf, RTAS_DATA_BUF_SIZE);
59 
60 	spin_unlock(&rtas_data_buf_lock);
61 	return rc;
62 }
63 
64 static int delete_dt_node(struct device_node *dn)
65 {
66 	pr_debug("removing node %pOFfp\n", dn);
67 	dlpar_detach_node(dn);
68 	return 0;
69 }
70 
71 static int update_dt_property(struct device_node *dn, struct property **prop,
72 			      const char *name, u32 vd, char *value)
73 {
74 	struct property *new_prop = *prop;
75 	int more = 0;
76 
77 	/* A negative 'vd' value indicates that only part of the new property
78 	 * value is contained in the buffer and we need to call
79 	 * ibm,update-properties again to get the rest of the value.
80 	 *
81 	 * A negative value is also the two's compliment of the actual value.
82 	 */
83 	if (vd & 0x80000000) {
84 		vd = ~vd + 1;
85 		more = 1;
86 	}
87 
88 	if (new_prop) {
89 		/* partial property fixup */
90 		char *new_data = kzalloc(new_prop->length + vd, GFP_KERNEL);
91 		if (!new_data)
92 			return -ENOMEM;
93 
94 		memcpy(new_data, new_prop->value, new_prop->length);
95 		memcpy(new_data + new_prop->length, value, vd);
96 
97 		kfree(new_prop->value);
98 		new_prop->value = new_data;
99 		new_prop->length += vd;
100 	} else {
101 		new_prop = kzalloc(sizeof(*new_prop), GFP_KERNEL);
102 		if (!new_prop)
103 			return -ENOMEM;
104 
105 		new_prop->name = kstrdup(name, GFP_KERNEL);
106 		if (!new_prop->name) {
107 			kfree(new_prop);
108 			return -ENOMEM;
109 		}
110 
111 		new_prop->length = vd;
112 		new_prop->value = kzalloc(new_prop->length, GFP_KERNEL);
113 		if (!new_prop->value) {
114 			kfree(new_prop->name);
115 			kfree(new_prop);
116 			return -ENOMEM;
117 		}
118 
119 		memcpy(new_prop->value, value, vd);
120 		*prop = new_prop;
121 	}
122 
123 	if (!more) {
124 		pr_debug("updating node %pOF property %s\n", dn, name);
125 		of_update_property(dn, new_prop);
126 		*prop = NULL;
127 	}
128 
129 	return 0;
130 }
131 
132 static int update_dt_node(struct device_node *dn, s32 scope)
133 {
134 	struct update_props_workarea *upwa;
135 	struct property *prop = NULL;
136 	int i, rc, rtas_rc;
137 	char *prop_data;
138 	char *rtas_buf;
139 	int update_properties_token;
140 	u32 nprops;
141 	u32 vd;
142 
143 	update_properties_token = rtas_token("ibm,update-properties");
144 	if (update_properties_token == RTAS_UNKNOWN_SERVICE)
145 		return -EINVAL;
146 
147 	rtas_buf = kzalloc(RTAS_DATA_BUF_SIZE, GFP_KERNEL);
148 	if (!rtas_buf)
149 		return -ENOMEM;
150 
151 	upwa = (struct update_props_workarea *)&rtas_buf[0];
152 	upwa->phandle = cpu_to_be32(dn->phandle);
153 
154 	do {
155 		rtas_rc = mobility_rtas_call(update_properties_token, rtas_buf,
156 					scope);
157 		if (rtas_rc < 0)
158 			break;
159 
160 		prop_data = rtas_buf + sizeof(*upwa);
161 		nprops = be32_to_cpu(upwa->nprops);
162 
163 		/* On the first call to ibm,update-properties for a node the
164 		 * the first property value descriptor contains an empty
165 		 * property name, the property value length encoded as u32,
166 		 * and the property value is the node path being updated.
167 		 */
168 		if (*prop_data == 0) {
169 			prop_data++;
170 			vd = be32_to_cpu(*(__be32 *)prop_data);
171 			prop_data += vd + sizeof(vd);
172 			nprops--;
173 		}
174 
175 		for (i = 0; i < nprops; i++) {
176 			char *prop_name;
177 
178 			prop_name = prop_data;
179 			prop_data += strlen(prop_name) + 1;
180 			vd = be32_to_cpu(*(__be32 *)prop_data);
181 			prop_data += sizeof(vd);
182 
183 			switch (vd) {
184 			case 0x00000000:
185 				/* name only property, nothing to do */
186 				break;
187 
188 			case 0x80000000:
189 				of_remove_property(dn, of_find_property(dn,
190 							prop_name, NULL));
191 				prop = NULL;
192 				break;
193 
194 			default:
195 				rc = update_dt_property(dn, &prop, prop_name,
196 							vd, prop_data);
197 				if (rc) {
198 					pr_err("updating %s property failed: %d\n",
199 					       prop_name, rc);
200 				}
201 
202 				prop_data += vd;
203 				break;
204 			}
205 
206 			cond_resched();
207 		}
208 
209 		cond_resched();
210 	} while (rtas_rc == 1);
211 
212 	kfree(rtas_buf);
213 	return 0;
214 }
215 
216 static int add_dt_node(struct device_node *parent_dn, __be32 drc_index)
217 {
218 	struct device_node *dn;
219 	int rc;
220 
221 	dn = dlpar_configure_connector(drc_index, parent_dn);
222 	if (!dn)
223 		return -ENOENT;
224 
225 	rc = dlpar_attach_node(dn, parent_dn);
226 	if (rc)
227 		dlpar_free_cc_nodes(dn);
228 
229 	pr_debug("added node %pOFfp\n", dn);
230 
231 	return rc;
232 }
233 
234 int pseries_devicetree_update(s32 scope)
235 {
236 	char *rtas_buf;
237 	__be32 *data;
238 	int update_nodes_token;
239 	int rc;
240 
241 	update_nodes_token = rtas_token("ibm,update-nodes");
242 	if (update_nodes_token == RTAS_UNKNOWN_SERVICE)
243 		return 0;
244 
245 	rtas_buf = kzalloc(RTAS_DATA_BUF_SIZE, GFP_KERNEL);
246 	if (!rtas_buf)
247 		return -ENOMEM;
248 
249 	do {
250 		rc = mobility_rtas_call(update_nodes_token, rtas_buf, scope);
251 		if (rc && rc != 1)
252 			break;
253 
254 		data = (__be32 *)rtas_buf + 4;
255 		while (be32_to_cpu(*data) & NODE_ACTION_MASK) {
256 			int i;
257 			u32 action = be32_to_cpu(*data) & NODE_ACTION_MASK;
258 			u32 node_count = be32_to_cpu(*data) & NODE_COUNT_MASK;
259 
260 			data++;
261 
262 			for (i = 0; i < node_count; i++) {
263 				struct device_node *np;
264 				__be32 phandle = *data++;
265 				__be32 drc_index;
266 
267 				np = of_find_node_by_phandle(be32_to_cpu(phandle));
268 				if (!np) {
269 					pr_warn("Failed lookup: phandle 0x%x for action 0x%x\n",
270 						be32_to_cpu(phandle), action);
271 					continue;
272 				}
273 
274 				switch (action) {
275 				case DELETE_DT_NODE:
276 					delete_dt_node(np);
277 					break;
278 				case UPDATE_DT_NODE:
279 					update_dt_node(np, scope);
280 					break;
281 				case ADD_DT_NODE:
282 					drc_index = *data++;
283 					add_dt_node(np, drc_index);
284 					break;
285 				}
286 
287 				of_node_put(np);
288 				cond_resched();
289 			}
290 		}
291 
292 		cond_resched();
293 	} while (rc == 1);
294 
295 	kfree(rtas_buf);
296 	return rc;
297 }
298 
299 void post_mobility_fixup(void)
300 {
301 	int rc;
302 
303 	rtas_activate_firmware();
304 
305 	/*
306 	 * We don't want CPUs to go online/offline while the device
307 	 * tree is being updated.
308 	 */
309 	cpus_read_lock();
310 
311 	/*
312 	 * It's common for the destination firmware to replace cache
313 	 * nodes.  Release all of the cacheinfo hierarchy's references
314 	 * before updating the device tree.
315 	 */
316 	cacheinfo_teardown();
317 
318 	rc = pseries_devicetree_update(MIGRATION_SCOPE);
319 	if (rc)
320 		pr_err("device tree update failed: %d\n", rc);
321 
322 	cacheinfo_rebuild();
323 
324 	cpus_read_unlock();
325 
326 	/* Possibly switch to a new L1 flush type */
327 	pseries_setup_security_mitigations();
328 
329 	/* Reinitialise system information for hv-24x7 */
330 	read_24x7_sys_info();
331 
332 	return;
333 }
334 
335 static int poll_vasi_state(u64 handle, unsigned long *res)
336 {
337 	unsigned long retbuf[PLPAR_HCALL_BUFSIZE];
338 	long hvrc;
339 	int ret;
340 
341 	hvrc = plpar_hcall(H_VASI_STATE, retbuf, handle);
342 	switch (hvrc) {
343 	case H_SUCCESS:
344 		ret = 0;
345 		*res = retbuf[0];
346 		break;
347 	case H_PARAMETER:
348 		ret = -EINVAL;
349 		break;
350 	case H_FUNCTION:
351 		ret = -EOPNOTSUPP;
352 		break;
353 	case H_HARDWARE:
354 	default:
355 		pr_err("unexpected H_VASI_STATE result %ld\n", hvrc);
356 		ret = -EIO;
357 		break;
358 	}
359 	return ret;
360 }
361 
362 static int wait_for_vasi_session_suspending(u64 handle)
363 {
364 	unsigned long state;
365 	int ret;
366 
367 	/*
368 	 * Wait for transition from H_VASI_ENABLED to
369 	 * H_VASI_SUSPENDING. Treat anything else as an error.
370 	 */
371 	while (true) {
372 		ret = poll_vasi_state(handle, &state);
373 
374 		if (ret != 0 || state == H_VASI_SUSPENDING) {
375 			break;
376 		} else if (state == H_VASI_ENABLED) {
377 			ssleep(1);
378 		} else {
379 			pr_err("unexpected H_VASI_STATE result %lu\n", state);
380 			ret = -EIO;
381 			break;
382 		}
383 	}
384 
385 	/*
386 	 * Proceed even if H_VASI_STATE is unavailable. If H_JOIN or
387 	 * ibm,suspend-me are also unimplemented, we'll recover then.
388 	 */
389 	if (ret == -EOPNOTSUPP)
390 		ret = 0;
391 
392 	return ret;
393 }
394 
395 static void prod_single(unsigned int target_cpu)
396 {
397 	long hvrc;
398 	int hwid;
399 
400 	hwid = get_hard_smp_processor_id(target_cpu);
401 	hvrc = plpar_hcall_norets(H_PROD, hwid);
402 	if (hvrc == H_SUCCESS)
403 		return;
404 	pr_err_ratelimited("H_PROD of CPU %u (hwid %d) error: %ld\n",
405 			   target_cpu, hwid, hvrc);
406 }
407 
408 static void prod_others(void)
409 {
410 	unsigned int cpu;
411 
412 	for_each_online_cpu(cpu) {
413 		if (cpu != smp_processor_id())
414 			prod_single(cpu);
415 	}
416 }
417 
418 static u16 clamp_slb_size(void)
419 {
420 	u16 prev = mmu_slb_size;
421 
422 	slb_set_size(SLB_MIN_SIZE);
423 
424 	return prev;
425 }
426 
427 static int do_suspend(void)
428 {
429 	u16 saved_slb_size;
430 	int status;
431 	int ret;
432 
433 	pr_info("calling ibm,suspend-me on CPU %i\n", smp_processor_id());
434 
435 	/*
436 	 * The destination processor model may have fewer SLB entries
437 	 * than the source. We reduce mmu_slb_size to a safe minimum
438 	 * before suspending in order to minimize the possibility of
439 	 * programming non-existent entries on the destination. If
440 	 * suspend fails, we restore it before returning. On success
441 	 * the OF reconfig path will update it from the new device
442 	 * tree after resuming on the destination.
443 	 */
444 	saved_slb_size = clamp_slb_size();
445 
446 	ret = rtas_ibm_suspend_me(&status);
447 	if (ret != 0) {
448 		pr_err("ibm,suspend-me error: %d\n", status);
449 		slb_set_size(saved_slb_size);
450 	}
451 
452 	return ret;
453 }
454 
455 /**
456  * struct pseries_suspend_info - State shared between CPUs for join/suspend.
457  * @counter: Threads are to increment this upon resuming from suspend
458  *           or if an error is received from H_JOIN. The thread which performs
459  *           the first increment (i.e. sets it to 1) is responsible for
460  *           waking the other threads.
461  * @done: False if join/suspend is in progress. True if the operation is
462  *        complete (successful or not).
463  */
464 struct pseries_suspend_info {
465 	atomic_t counter;
466 	bool done;
467 };
468 
469 static int do_join(void *arg)
470 {
471 	struct pseries_suspend_info *info = arg;
472 	atomic_t *counter = &info->counter;
473 	long hvrc;
474 	int ret;
475 
476 retry:
477 	/* Must ensure MSR.EE off for H_JOIN. */
478 	hard_irq_disable();
479 	hvrc = plpar_hcall_norets(H_JOIN);
480 
481 	switch (hvrc) {
482 	case H_CONTINUE:
483 		/*
484 		 * All other CPUs are offline or in H_JOIN. This CPU
485 		 * attempts the suspend.
486 		 */
487 		ret = do_suspend();
488 		break;
489 	case H_SUCCESS:
490 		/*
491 		 * The suspend is complete and this cpu has received a
492 		 * prod, or we've received a stray prod from unrelated
493 		 * code (e.g. paravirt spinlocks) and we need to join
494 		 * again.
495 		 *
496 		 * This barrier orders the return from H_JOIN above vs
497 		 * the load of info->done. It pairs with the barrier
498 		 * in the wakeup/prod path below.
499 		 */
500 		smp_mb();
501 		if (READ_ONCE(info->done) == false) {
502 			pr_info_ratelimited("premature return from H_JOIN on CPU %i, retrying",
503 					    smp_processor_id());
504 			goto retry;
505 		}
506 		ret = 0;
507 		break;
508 	case H_BAD_MODE:
509 	case H_HARDWARE:
510 	default:
511 		ret = -EIO;
512 		pr_err_ratelimited("H_JOIN error %ld on CPU %i\n",
513 				   hvrc, smp_processor_id());
514 		break;
515 	}
516 
517 	if (atomic_inc_return(counter) == 1) {
518 		pr_info("CPU %u waking all threads\n", smp_processor_id());
519 		WRITE_ONCE(info->done, true);
520 		/*
521 		 * This barrier orders the store to info->done vs subsequent
522 		 * H_PRODs to wake the other CPUs. It pairs with the barrier
523 		 * in the H_SUCCESS case above.
524 		 */
525 		smp_mb();
526 		prod_others();
527 	}
528 	/*
529 	 * Execution may have been suspended for several seconds, so
530 	 * reset the watchdog.
531 	 */
532 	touch_nmi_watchdog();
533 	return ret;
534 }
535 
536 /*
537  * Abort reason code byte 0. We use only the 'Migrating partition' value.
538  */
539 enum vasi_aborting_entity {
540 	ORCHESTRATOR        = 1,
541 	VSP_SOURCE          = 2,
542 	PARTITION_FIRMWARE  = 3,
543 	PLATFORM_FIRMWARE   = 4,
544 	VSP_TARGET          = 5,
545 	MIGRATING_PARTITION = 6,
546 };
547 
548 static void pseries_cancel_migration(u64 handle, int err)
549 {
550 	u32 reason_code;
551 	u32 detail;
552 	u8 entity;
553 	long hvrc;
554 
555 	entity = MIGRATING_PARTITION;
556 	detail = abs(err) & 0xffffff;
557 	reason_code = (entity << 24) | detail;
558 
559 	hvrc = plpar_hcall_norets(H_VASI_SIGNAL, handle,
560 				  H_VASI_SIGNAL_CANCEL, reason_code);
561 	if (hvrc)
562 		pr_err("H_VASI_SIGNAL error: %ld\n", hvrc);
563 }
564 
565 static int pseries_suspend(u64 handle)
566 {
567 	const unsigned int max_attempts = 5;
568 	unsigned int retry_interval_ms = 1;
569 	unsigned int attempt = 1;
570 	int ret;
571 
572 	while (true) {
573 		struct pseries_suspend_info info;
574 		unsigned long vasi_state;
575 		int vasi_err;
576 
577 		info = (struct pseries_suspend_info) {
578 			.counter = ATOMIC_INIT(0),
579 			.done = false,
580 		};
581 
582 		ret = stop_machine(do_join, &info, cpu_online_mask);
583 		if (ret == 0)
584 			break;
585 		/*
586 		 * Encountered an error. If the VASI stream is still
587 		 * in Suspending state, it's likely a transient
588 		 * condition related to some device in the partition
589 		 * and we can retry in the hope that the cause has
590 		 * cleared after some delay.
591 		 *
592 		 * A better design would allow drivers etc to prepare
593 		 * for the suspend and avoid conditions which prevent
594 		 * the suspend from succeeding. For now, we have this
595 		 * mitigation.
596 		 */
597 		pr_notice("Partition suspend attempt %u of %u error: %d\n",
598 			  attempt, max_attempts, ret);
599 
600 		if (attempt == max_attempts)
601 			break;
602 
603 		vasi_err = poll_vasi_state(handle, &vasi_state);
604 		if (vasi_err == 0) {
605 			if (vasi_state != H_VASI_SUSPENDING) {
606 				pr_notice("VASI state %lu after failed suspend\n",
607 					  vasi_state);
608 				break;
609 			}
610 		} else if (vasi_err != -EOPNOTSUPP) {
611 			pr_err("VASI state poll error: %d", vasi_err);
612 			break;
613 		}
614 
615 		pr_notice("Will retry partition suspend after %u ms\n",
616 			  retry_interval_ms);
617 
618 		msleep(retry_interval_ms);
619 		retry_interval_ms *= 10;
620 		attempt++;
621 	}
622 
623 	return ret;
624 }
625 
626 static int pseries_migrate_partition(u64 handle)
627 {
628 	int ret;
629 
630 	ret = wait_for_vasi_session_suspending(handle);
631 	if (ret)
632 		return ret;
633 
634 	ret = pseries_suspend(handle);
635 	if (ret == 0)
636 		post_mobility_fixup();
637 	else
638 		pseries_cancel_migration(handle, ret);
639 
640 	return ret;
641 }
642 
643 int rtas_syscall_dispatch_ibm_suspend_me(u64 handle)
644 {
645 	return pseries_migrate_partition(handle);
646 }
647 
648 static ssize_t migration_store(struct class *class,
649 			       struct class_attribute *attr, const char *buf,
650 			       size_t count)
651 {
652 	u64 streamid;
653 	int rc;
654 
655 	rc = kstrtou64(buf, 0, &streamid);
656 	if (rc)
657 		return rc;
658 
659 	rc = pseries_migrate_partition(streamid);
660 	if (rc)
661 		return rc;
662 
663 	return count;
664 }
665 
666 /*
667  * Used by drmgr to determine the kernel behavior of the migration interface.
668  *
669  * Version 1: Performs all PAPR requirements for migration including
670  *	firmware activation and device tree update.
671  */
672 #define MIGRATION_API_VERSION	1
673 
674 static CLASS_ATTR_WO(migration);
675 static CLASS_ATTR_STRING(api_version, 0444, __stringify(MIGRATION_API_VERSION));
676 
677 static int __init mobility_sysfs_init(void)
678 {
679 	int rc;
680 
681 	mobility_kobj = kobject_create_and_add("mobility", kernel_kobj);
682 	if (!mobility_kobj)
683 		return -ENOMEM;
684 
685 	rc = sysfs_create_file(mobility_kobj, &class_attr_migration.attr);
686 	if (rc)
687 		pr_err("unable to create migration sysfs file (%d)\n", rc);
688 
689 	rc = sysfs_create_file(mobility_kobj, &class_attr_api_version.attr.attr);
690 	if (rc)
691 		pr_err("unable to create api_version sysfs file (%d)\n", rc);
692 
693 	return 0;
694 }
695 machine_device_initcall(pseries, mobility_sysfs_init);
696